Classification

The vehicle classification process is not an easy problem. Truck vehicles are very similar to each other and moreover they contain a lot of confusing information (like company name, driver name, etc.). Images of trucks also have a signifficant shift in position and they were taken under difficult weather conditions like snowing or during night and with lights on. Images of cars have a lot of reflections from metal surface and quite big variance within class.

Statistical pattern recognition approaches for pattern detection generally fall into two categories: generative methods or discriminative methods depending on the estimation criteria used for adjusting the model parameters and/or structure.

Generative approaches such as hidden Markov Model, Markov random field or naive Bayes classifier relies on probability distribution estimation over examples using maximum likelihood or maximum a posterior methods. Discriminative methods such as neural networks, LDA or SVM are trying to find a decision boundary between classes. Since generative mixture methods such as a mixture of Gaussians and a mixture of factor analyzers rely on a joint probability distribution over examples it carries some advantages like handling incomplete examples, the typical estimation criterion is nevertheless optimal from the classification viewpoint. Furthermore generative approaches usually require larger data set in comparison to discriminative methods since most of them are focusing on covariance matrix estimation. Discriminative methods that focus directly on the parametric decision boundary, e.g. Fisher's linear discriminant(FLD) or SVM, typically yield a better classification results[28].

We decied to try these classification algorithms:

• k-NN (Nearest Neighbor)
• Fisher's Linear Discriminant (FLD)

The first one, k-NN (Nearest Neighbor), serves us as a baseline, it is easy to implement and the results are quite good. We will tune the SIFT parameters using this algorithm. We will also show how our different SIFT representations can affect the classification rank. Fisher's Linear Discriminant is a robust method with quite a small amount of parameters which usually gives good results.

We investigated two distance measures to compare testing and training samples, the Euclidean Euclidean distance and additionally also Earth Mover's Distance (EMD).